Learn more, waste less

Detect waste with AI in Gdańsk

Detect waste project

Detect waste project

WiMLDS Project Trójmiasto - AI4Good

WiMLDS Gdańsk

Detect waste with AI

Detect waste is a non-profit, educational, eco project that aims to use Artificial Intelligence for the general good. We have gathered a team of ten carefully selected members and five mentors to work together on the problem of the world’s waste pollution.

For over 5 months (October 2020 - March 2021) we learn and work together on developing a model for trash detection. The project is aimed at people from Pomerania, with at least intermediate Python knowledge and an experience in Deep Learning.

Waste pollution is one of the biggest environmental issues in the modern world. The massive production of disposable goods in the last years resulted in an exponential increase in produced waste. It’s frightening how much waste we produce – it’s about 2 billion tons every year. In these times, every single thing that we buy must be covered with some wrapping. Sometimes even peeled fruits are put into a plastic bag.

Dates

What & when

 
 
 
 
 

Recruting members

Online - apply here

08 Oct 2020 – 18 Oct 2020 Gdańsk, Poland
 
 
 
 
 

Participants selection

18 Oct 2020 – 20 Oct 2020
Summary: We have selected 7 outstanding women from various backgrounds to join our project.
 
 
 
 
 

Project launch

27 Oct 2020 – 03 Nov 2020 Gdańsk, Poland
Summary: First meeting and first sprint ended successfully.
 
 
 
 
 

Collecting & analyzing data

03 Nov 2020 – 24 Nov 2020 Gdańsk, Poland
Goal: Getting to know TACO dataset, statistical analyzes. Summary: We have performed EDA with multiple statistics. It allowed u to plan the next sprint. We have decided about waste categorization: we follow Gdańsk recycling rules.
 
 
 
 
 

Model design & training

24 Nov 2020 – 22 Dec 2020 Gdańsk, Poland
Goal: Literature review, basic models and common solutions. Summary: We have selected a few detection models to check out and learn about: YOLO v4, DETR, EfficientDets. Additionaly we have checked one segmentation model: Masked RCNN. We have recived 3000 more annotated waste images form Epinote. We divided work into teams, so we can compare final results. Every team uploaded their work to our Github.
 
 
 
 
 

Model design, training & testing

22 Dec 2020 – 19 Jan 2021 Gdańsk, Poland
Goal: Going deeper - advanced models and data augmentation. Summary: We have decided to focus only on EfficientDets which gave us the best results. Additionally we have started to try out some data augmentation techniques: classic transformations as zoom in/out, rotation but also color augmentation and artificial snow and rain. Moreover we have discovered multiple more datasets with waste designed for segmentation, detection and classification which we added to our TACO dataset. Huge boost of data made us to re-define models’ pipeline.
 
 
 
 
 

Final model selection

19 Jan 2021 – 02 Feb 2021 Gdańsk, Poland
Goal: Finishing the project. Summary: We have noticed that our models works drastically better when classifying only single class: litter. Moreover, we have over 100k images of litter for classification. Our final pipeline is to detect litter with EfficientDet model and classify detected waste into 7 categories with EfficientNet.
 
 
 
 
 

Summary

Finishing the project

02 Feb 2021 – 27 Mar 2021 Gdańsk, Poland
After five months and around 1000 (wo)man-hours of uphill battle we successfully finished our project. The “detect waste” will be a memorable experience for many of us. We finished with over 80 commits to the official repository, 500 experiments in 5 projects on neptune.ai (and untracked numerous trainings), thousands of lines of code. Click the link the summary.

Plastic is everywhere

Motivation

What and why?

Today, more than 300 million tons of plastic are produced annually. Plastic is everywhere and we constantly use it in our daily life.

The idea of Detect waste in Pomerania is to use Artificial Intelligence to detect various waste in the environment. Our solution is applicable for video and photography.

Our goal is to use AI for Good.

Contact

Plastic Object Detection - AI4Good